Robust Estimation of Stochastic Data in Machine Learning Systems

The instability of algorithms for filtering input data, usually stochastic, to the a priori uncertainty of their probabilistic characteristics is a serious problem of machine learning. Since such uncertainty occurs in a huge number of cases of developing and applying of artificial intelligence systems (in particular, when using streaming learning, preparing data for training, and evaluating models), the problem of ensuring accuracy and stability of algorithms for evaluation of stochastic data with unknown statistical characteristics for machine learning systems is one of the most urgent. The existing methods of nonlinear filtering in a general formulation do not allow to solve this problem. In this connection, the paper proposes a method of robust stochastic filtering, oriented to the processing of discrete data in machine learning systems. The method allows to perform dynamic estimation of a nonlinear discrete stochastic vector of state perturbed by correlated noise with an unknown probability distribution belonging to the class of distributions with limited second moments (mean squares). At the same time, the observation of this vector is also carried out in conditions of interference with unknown probabilistic characteristics. The robust estimation algorithm is synthesized by minimizing a new criterion that depends on the nonlinear measurement residual function determined by the class of the measurement noise distribution function and takes into account the correlation of state vector disturbances. In contrast to the currently developed nonlinear filtering algorithms, the proposed method does not involve knowledge of the laws of probability distributions of state vector disturbances and measurement interference, but only their belonging to a certain class of distributions At the same time, its computational implementation requires significantly less computations compared to the known nonlinear filtering methods due to the coincidence of the dimensionality of the obtained filter with the dimensionality of the estimated state vector. Such advantageous features of the developed method provide the possibility of its effective practical application in processing under uncertainty of input data of artificial intelligence systems used in various technical systems - information and control, measurement, info-communication, navigation, avionics systems and others.